Teaching AI to Self-Diagnose: Probing Hidden States of Large Language Models via a Questioning Mechanism

Original Author & Source:

• Original Author/Maintainer: arXiv authors
• Source Platform: arXiv
• Original Title: What Am I Missing? Question-Answering as Hidden State Probing
• Original Link: http://arxiv.org/abs/2605.31561v1
• Source Publication/Update Time: 2026-05-29T17:27:07Z

Core Introduction: The study proposes an innovative "Student-Teacher" framework that allows large language models to diagnose uncertainties in their reasoning process through self-questioning. By analyzing the hidden state signals generated when the model formulates questions, the correctness of the final answer can be predicted, providing a new perspective on the self-correction capabilities of large models. The study finds that the model has strong self-diagnosis ability but weak correction ability, and questioning intervention has a double-edged sword effect.

Research Background: The Uncertainty Challenge in Reasoning Processes

Since the introduction of Chain-of-Thought technology into large language models, inference during testing has become an important research direction. However, a long-standing problem plaguing researchers is that even with the same input prompts or even intermediate steps, multiple samplings of the model still produce different answers.

This uncertainty exposes a core blind spot in the reasoning mechanism—lack of in-depth understanding of the model's "thinking process". Traditional methods rely on final outputs to evaluate reasoning quality, ignoring the rich information in internal hidden states.

Core Innovation: Questioning as a Tool for Probing Hidden States

This paper proposes a disruptive idea: using "questioning" as an intervention method during reasoning to reveal the model's hidden states. A "Student-Teacher" framework is designed: the student model asks questions to the teacher model, and researchers train a probing model to analyze the hidden states of the student before and after asking questions.

Key Finding: The probing model can predict whether the final reasoning trajectory is correct before the teacher answers. This indicates that the self-diagnosis of the model when formulating questions is more valuable than the teacher's information—the model exposes its uncertainty when clarifying its confusion.

Technical Implementation: Design of the Gating Strategy

Based on the findings, the study formalizes the questioning behavior as a sequential decision-making problem, using the quality score output by the probing model to define a gating strategy that determines when to ask questions to maximize the probability of correct answers. The core logic of the strategy:

1. Real-time Monitoring: Continuously monitor changes in the model's hidden states
2. Uncertainty Quantification: Evaluate the reliability of the reasoning path through the probing model
3. Selective Intervention: Trigger questions only when uncertainty is high
4. Dynamic Adjustment: Optimize the questioning strategy based on feedback

Key Finding: The Gap Between Diagnosis and Correction

Experimental results show that the success of questioning intervention depends on the model's self-consistency. There is an obvious "gap"—the gating strategy can effectively identify correctness and uncertainty, but the probability of destroying correct trajectories is equivalent to the probability of repairing wrong trajectories.

Specific Performance:

• Strong Detection Ability: Accurately identify its own uncertain states
• Weak Correction Ability: Identifying uncertainty does not automatically lead to effective resolution
• Double-Edged Sword Effect: Questioning can both save wrong reasoning and disrupt correct thinking

This raises questions about the self-correction ability of large models: merely "realizing mistakes" is not enough; more refined correction mechanisms are needed.

Practical Significance and Future Outlook

Immediate Application Value

1. Reasoning Quality Evaluation: Predict answer quality without complete output
2. Dynamic Computing Allocation: Increase reasoning depth when uncertain, reduce overhead when certain
3. Human-Machine Collaboration Optimization: Identify moments when the model needs help to improve interaction efficiency

Long-Term Research Directions

1. Bridging the Diagnosis-Correction Gap: Develop algorithms that can identify and effectively correct errors
2. Multimodal Expansion: Apply hidden state probing to tasks such as vision and code generation
3. Model Architecture Improvement: Design model structures that inherently have better self-diagnosis capabilities

Conclusion

This study reminds us that the "black box" nature of large language models is not only reflected in the difficulty of understanding outputs but also in the difficulty of grasping the "thinking process". By redefining questioning as a metacognitive tool, researchers have provided a window into the inner workings of the model.

Although the gap between diagnosis and correction indicates that there is still a long way to go, this is exactly the charm of scientific exploration—each discovery leads to new questions, and each breakthrough opens up new possibilities.